Multivariate Feature Extraction
Xiangxiang Xu, Lizhong Zheng · 2022
We propose a general framework to extract features for learning problems involving multiple variables. First, we decompose multivariate dependence into different components to obtain the piece relevant to the learning task. Then, we establish a modal decomposition approach to represent the component as informative features. We further demonstrate the algorithm design for extracting such features from real data, which can incorporate and utilize existing deep feature extractors. We also present an application of our framework in learning tasks with side information.